EN: A fix is ready on a fork branch (see links below), but this repository currently rejects pull request creation from my account (hezhongtang does not have the correct permissions to execute CreatePullRequest), so I am opening this tracking issue instead. A maintainer can either grant PR permission or push the branch and open the PR directly — the full bilingual PR description is included below.

中文： 修复已在 fork 分支上就绪（链接见下），但仓库当前拒绝我的账号创建 PR（hezhongtang does not have the correct permissions to execute CreatePullRequest），因此先开此跟踪 issue。维护者可以授予 PR 权限，或直接推送该分支并创建 PR——完整的双语 PR 描述附在下方。

• Compare / 对比: main...hezhongtang:minimax-code:fix/k3-reasoning-effort
• Branch / 分支: hezhongtang:fix/k3-reasoning-effort (content identical to hezhongtang/minimax-code@main)

fix(model-system): import reasoning effort options from models.dev presets

Summary / 概述

EN: The provider preset parser only reads the boolean reasoning flag from models.dev and drops the reasoning_options metadata. Models that declare configurable reasoning effort — e.g. Moonshot kimi-k3 (moonshotai / moonshotai-cn, reasoning_options: [{type: 'effort', values: ['low','high','max']}]) — are imported as reasoning: true only, so:

• the TUI effort selector shows no levels for K3
• mcode exec --effort high is rejected by local validation
• the existing openai-completions → reasoning_effort wire mapping can never be reached through the normal user path

中文： 目前的 provider preset 解析器只读取 models.dev 的布尔字段 reasoning，丢弃了 reasoning_options 元数据。对于声明了可配置推理强度（reasoning effort）的模型——例如月之暗面 kimi-k3（moonshotai / moonshotai-cn，reasoning_options: [{type: 'effort', values: ['low','high','max']}]）——导入时只保留了 reasoning: true，导致：

• TUI 的 effort 选择器中 K3 没有任何可选档位
• mcode exec --effort high 被本地校验拒绝
• 已有的 openai-completions → reasoning_effort 请求映射无法通过正常用户路径触达

Fix / 修复方案

EN: Add a generic parseModelEffortOptions step to the preset parser that normalizes the {type: 'effort', values: [...]} entry of reasoning_options into UserModelInputView.effortOptions (trim, drop empty/non-string values, dedupe, preserve declared order; no output when empty).

• No hard-coded K3 / provider special-casing — any models.dev model declaring effort values benefits automatically (both Moonshot providers included).
• Toggle-only entries and bare reasoning: true still produce no effortOptions; reasoning is not generalized into effort support.
• No changes to TUI, headless validation, model-ref, or transport: the existing persistence (thinking.effortOptions), roster exposure, --effort validation, and reasoning_effort request patch flow through unchanged. MiniMax M3 on/off logic untouched.

中文： 在 preset 解析器中新增通用的 parseModelEffortOptions 步骤，将 reasoning_options 中的 {type: 'effort', values: [...]} 条目规范化到 UserModelInputView.effortOptions（去除首尾空白、丢弃空值与非字符串、去重、保持声明顺序；结果为空则不输出）。

• 不针对 K3 或特定 provider 做硬编码——任何在 models.dev 声明 effort 档位的模型都会自动受益（两个 Moonshot provider 均包含）。
• 仅有 toggle 的条目和裸 reasoning: true 仍然不会产生 effortOptions；不会把"支持思考"泛化为"支持 effort 选择"。
• 不改动 TUI、headless 校验、model-ref 和传输层：已有的持久化（thinking.effortOptions）、模型列表暴露、--effort 校验以及 reasoning_effort 请求补丁流程均原样复用。MiniMax M3 的开关逻辑不受影响。

Tests / 测试

EN:

• Preset parser: K3 fixtures for both moonshotai and moonshotai-cn parse ['low','high','max']; toggle-only / empty-options models do not gain effortOptions; normalization keeps unknown values, dedupes, drops empties.
• Persistence: K3 input saves { reasoning: true, thinking: { effortOptions: [...] } }.
• Roster: list-models exposes effortOptions for K3.
• Headless: --effort low|high|max accepted; --effort medium rejected with Available levels: low, high, max.
• End-to-end: config metadata → modelRef → thinking protocol → requestPatch { reasoning_effort } for all three levels.
• CLI e2e against a mock OpenAI-compatible transport: --effort low/high/max produce wire reasoning_effort of low/high/max; medium is rejected locally before any request is sent.

中文：

• Preset 解析器：moonshotai 与 moonshotai-cn 的 K3 fixture 均能解析出 ['low','high','max']；仅 toggle / 空 options 的模型不会获得 effortOptions；规范化保留未知档位、去重、丢弃空值。
• 持久化：K3 输入保存为 { reasoning: true, thinking: { effortOptions: [...] } }。
• 模型列表：list-models 为 K3 暴露 effortOptions。
• Headless：--effort low|high|max 均被接受；--effort medium 被拒绝并提示 Available levels: low, high, max.
• 端到端：配置元数据 → modelRef → thinking protocol → 三个档位的 requestPatch { reasoning_effort }。
• CLI 端到端（模拟 OpenAI 兼容传输）：--effort low/high/max 产生线上的 reasoning_effort 为 low/high/max；medium 在发送任何请求前即被本地拒绝。

Verification / 验证

EN:

• Targeted vitest files: 154/154 passing
• pnpm typecheck, pnpm check:source, pnpm build, node --test test/byok.test.mjs (11/11) passing
• 12 failures in tui-app.test.ts / update-application.test.ts are pre-existing on the base commit (reproduced identically on base commit 81b666a; locale-dependent assertions) and unrelated to this change
• The reasoning_options fixtures match the live models.dev catalog (moonshotai/moonshotai-cn kimi-k3)

中文：

• 相关 vitest 文件：154/154 通过
• pnpm typecheck、pnpm check:source、pnpm build、node --test test/byok.test.mjs（11/11）均通过
• tui-app.test.ts / update-application.test.ts 中的 12 个失败为基线提交上已存在的问题（已在基线提交 81b666a 上以完全相同的方式复现；为依赖语言环境的断言），与本改动无关
• reasoning_options 测试 fixture 与线上 models.dev 目录一致（moonshotai/moonshotai-cn 的 kimi-k3）